Research On Defect Classification Of Diffuse Reflective Surface Based On Deep Learning

With the development of industrial modernization,the traditional human eye detection method has been unable to meet the daily production of industrial products.In the recent years,under the call of the national policy "machine replacement",automatic optical detection technology based on machine vision has gradually replaced human eye detection and has become the mainstream detection method.However,as people's pursuit of product appearance continues to improve,product surface properties have become diversified,such as the smooth metal surface,matte painted surface,and plastic painted surface of tablets or mobile phone.In a diffuse reflective surface like the matte paint surface,the surface particles can cause irregular diffuse reflection of the detection light,making defect information easily covered by complex background scattered light,so that the defect is difficult to be detected.Therefore,in order to solve the above problems,this paper conducts research on defect detection of diffuse reflection surfaces.The main contents include:Aiming at the problem that diffuse reflection surface defects are difficult to image,this paper analyzed the characteristics of the Micro Scattering Dark Filed(MSDF)image system in the inspection of ultra-smooth surface defects and the surface reflection characteristics of diffuse reflective surface,and designs the Coaxial Parallel Light(CPL)image system,which realizes the image of diffuse reflective surface defects.In view of the fact that the surface of the product may have both a diffuse reflection surface and a smooth surface,a bright field imaging system is designed to detect the dent defection on the smooth surface.Aiming at the problem that traditional image classification algorithms have low accuracy in classifying diffuse reflection defects,and the algorithm design process is cumbersome and lacks robustness,this paper propose to use convolutional neural networks in deep learning to classify defect images.In order to meet the requirements for rapid detection of defects at the factory production site,this paper first uses the two-threshold-based auto-correlation template matching method to realize the localization and segmentation of defect images,and then sends the segmented defect images to a convolutional neural network to classify the defect images,which realizes the rapid detection of defects.In order to train the convolutional neural network model,an image data set with 8 defect types is established,and the data set is expanded by date enhancement.In order to propose a convolutional neural network model suitable for defect image classification,this paper tests three classic convolutional neural network models with different characteristics,which is AlexNet,ResNet and GoogLeNet.The experiment show that the GoogLeNet model with the Inception structure has the highest average accuracy rate,reaching 90.5%,but its detection time is as long as 27ms,which cannot meet the needs of fast detection.Therefore,based on the characteristics of single channel and small resolution of defect image,and the design principle of balance the width and depth of network,this paper optimized the structure of the GoogLeNet model,and finally successfully improved the average accuracy of network model to 91.7%,and shortened the detection time to 12.67ms.